An agent system built around OpenAI's internal Codex. A person defines the task; coding agents implement it and respond to CI results; specialist agents review changes under risk-based routing. A human approves production deployment before an agent manages the rollout and monitors its results.
Human-defined change through green pull request, agent review, risk routing, human-approved rollout; and the production alert to proposed perf fix.
Implement and babysit the pull request
Codex makes code changes toward the goal, builds and tests, and babysits the pull request by fixing CI failures until it is green[Source 1]
Per-change deploy agent
After human approval for production, an agent manages the rollout, builds a monitoring dashboard, and watches signals of success or failure[Source 1]
Perf Factory
Agents sift through alerts and dashboards, de-duplicate signals, identify real latency regressions, root-cause them, and propose fixes[Source 1]
Where people stay involved
People review, approve, or guide the agent’s work here.
Reported: A human approves production deployment before the deploy agent takes over; opted-in low-risk areas auto-approve with review by exception. The Perf Factory boundary is the unknown catalog judgment kept in research details.
Supervision level
Exception-only
A person returns when the system raises an exception.
5/5Level 5 of 5
low-risk pull request -> merge in opted-in codebase areas[Source 1]
Supervision level
Work-product review
A person reviews the draft or implementation.
3/5Level 3 of 5
code change -> production rollout on the general path[Source 1]
Supervision level
Not classified
production alert -> proposed performance fix[Source 1]
Not reportedThe article names no underlying model for the internal harness. The harness-engineering team names gpt-5 and codex-cli for its own 2025 prototype start, which does not state the current internal harness model.
Harness
Internal Codex, described as much more advanced than the external product because it is plugged into almost every OpenAI system; ChatGPT Work runs on the same harness[Source 1, Source 5, Source 7]
Sandbox
Not reportedThe pipeline article documents no execution sandbox. The Agents API post documents sandboxing for the external product, not this internal pipeline boundary.
Tool access
Git repositories and GitHub, Slack, Notion, Databricks, Datadog, and internal logs and data sources[Source 1]
Knowledge
OpenAI moved its documentation inside the source code; internal Codex skills, some maintained by Codex itself; new engineers are directed to ask Codex during onboarding[Source 1, Source 3, Source 9]
Context management
The /goal setting lets an agent work until a goal is complete; threads run for days and spin off other agents[Source 1]
Credentials
Not reportedCredential handling for the internal system access is not documented; the production approval gate is an authorization step cited with human involvement.
Changes are classified by risk; high-risk changes can trigger more agent reviews or a mandated human review, while opted-in low-risk areas use an agent that auto-approves pull requests[Source 1]
Validation and failure handling
Domain-specialist review agents
Multiple agents, each configured as a domain specialist, review every change; the article compares this to a review by a domain expert from each relevant infrastructure team[Source 1, Source 2, Source 3]
Perf harness with Synthetics A/B
A perf harness sends problematic pull requests to the Synthetics A/B framework to evaluate performance implications[Source 1]
Reported observations
These results come from the sources. Notes below each result explain what it covers and any reported limits.
Headline claim
Almost all OpenAI employees used Codex and ChatGPT Work weekly as of the September 2026 report (self-reported).[Source 1, Source 8]
Reported by
OpenAI
Scope
Weekly Codex and ChatGPT Work use across OpenAI employees
Method
Internal token-usage tracking by department, chart sourced to OpenAI
The source does not report the denominator of this figure.
Observed in September 2026
About this result
Category
Adoption output
Evidence type
Reported measurement
What it covers
Weekly Codex and ChatGPT Work use across OpenAI employees, September 2026
Key observation
Non-engineering orgs such as finance, recruitment, and legal went from about 0% to 90% Codex usage within a four-month period (self-reported)[Source 1]
Reported by
OpenAI
Scope
Weekly Codex use in non-engineering orgs such as finance, recruitment, and legal
The source does not report the denominator of this figure.
Observed in September 2026
About this result
Category
Adoption output
Evidence type
Reported measurement
What it covers
Weekly Codex use in non-engineering orgs such as finance, recruitment, and legal over a four-month period
Key observation
Some build-test-deploy systems saw about a 10x load increase within roughly six months (self-reported)[Source 1]
Reported by
OpenAI
Scope
Load on build-test-deploy pipeline systems
The source does not report the denominator of this figure.
Observed in September 2026
About this result
Category
Runtime capacity
Evidence type
Estimate
What it covers
Load on build-test-deploy pipeline systems over about six months
Key observation
92% of OpenAI technical staff used Codex daily, as reported at DevDay in October 2025[Source 2]
Reported by
OpenAI
Scope
Daily Codex use by OpenAI technical staff in October 2025
The source does not report the denominator of this figure.
Observed in October 2025
About this result
Category
Adoption output
Evidence type
Reported measurement
What it covers
Daily Codex use by OpenAI technical staff in October 2025
Key observation
Engineers using Codex completed 70% more pull requests each week, as reported at DevDay in October 2025[Source 2]
Reported by
OpenAI
Scope
Weekly pull requests completed by OpenAI engineers using Codex, relative to unstated baseline
The source does not report the denominator of this figure.
Observed in October 2025
About this result
Category
Effectiveness
Evidence type
Reported measurement
What it covers
Weekly pull requests completed per OpenAI engineer using Codex, October 2025, against an unstated baseline
Lessons
Reported opinion: OpenAI's desktop lead attributes some adoption growth to longer-running tasks and an agent that coordinates sub-agents.[Source 1]
Reported opinion: OpenAI's desktop lead says role-specific and team-specific plugins helped people adapt agents to their work.[Source 1]
Reported: The article documents the pipeline as agent runs: implement and keep the PR green, handhold the approved rollout, and the production perf-fix loop. The review agents and the perf harness are cited with validation; the risk routing is a mechanism.
human involvement
Reported: A human approves production deployment before the deploy agent takes over; opted-in low-risk areas auto-approve with review by exception. The Perf Factory boundary is the unknown catalog judgment kept in research details.
implementation
Reported
validation
Reported: Domain-specialist agents review every change and the perf harness sends problematic PRs to Synthetics A/B; the CI suite runs inside the implement primitive.
Company usage claim relayed by an independent reporter; no method or denominator published. OpenAI's own 2026-09-10 launch post states nearly 100% of teams inside OpenAI, including finance and sales, use ChatGPT Work and Codex (a teams denominator, not employees).
Reported by
OpenAI
Scope
Weekly Codex and ChatGPT Work use across OpenAI employees
Denominator
Not reported
Method
Internal token-usage tracking by department, chart sourced to OpenAI
SupportsChatGPT is now a partner for your most ambitious workPreserved content.md, line 34 (OpenAI's own launch post: nearly 100% of teams inside OpenAI, including finance and sales, use ChatGPT Work and Codex; teams denominator, OpenAI self-report)
No explicit confidence assessment is recorded for this claim.
SupportsInside OpenAI's agentic software factoryPreserved content.md, line 48 (internal Codex is a lot more advanced than its external counterpart; ChatGPT Work is powered by the Codex harness)
SupportsIntroducing the Agents APIPreserved content.md, lines 12-14 (OpenAI scaled Codex and ChatGPT for Work on the same harness; the Agents API exposes 'that same harness and infrastructure that powers Codex')
SupportsHarness engineering: leveraging Codex in an agent-first worldPreserved content.md, lines 30, 73, and 117 (structured docs/ directory as the system of record; a roughly 100-line AGENTS.md serving as a map; a recurring doc-gardening agent; the initial AGENTS.md was written by Codex)
SupportsTranscript: 'How OpenAI's Codex Team Uses Their Coding Agent'Preserved content.md, lines 236, 258, and 274 (automations that keep PRs mergeable by resolving merge conflicts, a random-file bug-finder that runs multiple times a day, and a bot that quietly fixes bugs in recently merged PRs)
SupportsHarness engineering: leveraging Codex in an agent-first worldPreserved content.md, lines 46 and 48 (Codex requests additional specific agent reviews locally and in the cloud until all agent reviewers are satisfied; 'almost all review effort' handled agent-to-agent; one repository)
Contextualizeszbrock: many internal teams adopted the same practicesPreserved content.md, lines 12 and 16 (harness-engineering co-author: many internal teams adopted the same practices; some run centralized agent-mediated integration queues and local Codex threads that monitor CI; no self-built deploy dashboards described)
Company adoption figures relayed by an independent reporter; the four-month window carries no calendar dates. OpenAI's launch post corroborates finance-org use (month-end close reduced from days to hours) but not the 0%-to-90% trajectory.
Reported by
OpenAI
Scope
Weekly Codex use in non-engineering orgs such as finance, recruitment, and legal
ContextualizesChatGPT is now a partner for your most ambitious workPreserved content.md, line 37 (OpenAI's finance org uses ChatGPT Work for month-end close and forecasting, reduced from days to hours; corroborates non-engineering adoption, not the 0%-to-90% trajectory)
Andrew Ambrosino offers this explanation in the reported interview. The adoption timeline does not isolate the effect of longer runs from other product changes.
The article documents agent auto-approve for low-risk pull requests in opted-in areas and risk classification that routes high-risk changes to stricter review; the run-time exception path is not described. The harness-engineering post corroborates agent automerge of low-friction pull requests in one repository but describes no risk tiering, so the routing mechanism itself stays single-sourced.
Observation date
2026-09
SupportsInside OpenAI's agentic software factoryPreserved content.md, line 159 (risk classification; low-risk areas can opt in to an agent that auto-approves pull requests)
Pipeline step 6 states that a human approves a change to go to production before the deploy agent takes over.
Observation date
2026-09
SupportsInside OpenAI's agentic software factoryPreserved content.md, lines 165-175 (after a human approves production, an assigned agent handholds the rollout)
The article documents Perf Factory only up to agents that de-duplicate signals, identify real latency regressions, root-cause them, and propose fixes; whether a human reviews, applies, or lands those proposals is not stated.
